The answer to high-performance AI: in-situ photonic accelerator

Artificial intelligence (AI) has been developed to overthrow the traditional industrial thought through fast learning and inference, which has certainly made our lives easier and more efficient. Artificial neural network (ANN), as a representative of machine learning algorithms, has the widest range of applications solving problems of real world. However, ANN deployed in real world is limited by low computation speed, high energy consumption and few computing parallelism, despite nowadays being accelerated by Graphics Processing Unit (GPU), Application-Specific Integrated Circuit (ASIC) and Field-Programmable Gate Array (FPGA) chips. To address this challenge, “Photonic neuromorphic computation appears set to offer huge opportunities”, raised by one tutorial “Rise of the learning machines” from Nature Photonics. Nevertheless, one major challenge in successful photonics integration is the implementation of efficient nonlinear activation functions as stressed by Professor Aydogan Ozcan, Chancellor's Professor at University of California, Los Angeles (UCLA), in his Q&A titled “Machine learning with light”:

“Challenges do exist for all-optical implementations due to lack of highly efficient and practical nonlinear optical processes that can serve as activation functions for large numbers of nodes.”

Hence, search of an efficient solution for photonic nonlinear activation functions has spawned much research activities. In a new paper published in Light Science & Application, a team of scientists, led by Professor Aaron Voon-Yew Thean, Dean of College of Design & Engineering at National University of Singapore (NUS), from Department of Electrical and Computer Engineering, National University of Singapore, Singapore, has developed a novel photonic neural network accelerator based on a non-volatile Opto-Resistive RAM Switch to achieve programmable nonlinear activation functions. The photonic neural network accelerator adopts an optical-electrical hybrid architecture, consisting of a Mach–Zehnder interferometer with phase shifters in both arms, a coupler, a novel Opto-Resistive RAM Switch, and an electrical control unit. Interestingly, the MoS₂-based Opto-Resistive RAM Switch, the key component of the accelerator, is first designed and optimized to achieve excellent linearity between input optical power and switching voltage. Additionally, overcoming the restricted linear response in typical photonic components such as traditional photodetector, the Opto-Resistive RAM Switch can be used owing to its nonlinear switching. It is also worth noting that this is the first-time two-dimensional material is used in photonic neural network. Coupled with carefully tuned phase shifters and electrical control unit, the accelerator achieves superior programmability which can be cleverly used to accurately represent different nonlinear activation functions, including sigmoid, softplus, and clamped Rectified Linear Unit (ReLU). These are extremely critical in the acceleration of ANN.

The accelerator has been demonstrated in a complete in-situ photonic neural network on MNIST handwritten digit recognition task. Using the newly developed accelerator, the team have achieved a substantial accuracy of 91.6%, significantly reduced the average power consumption by 20 times and shrunk the hardware footprint by close to 40% when compared with other state-of-the-art architectures. With ultrahigh power efficiency and computation density, the reported architecture opens up new avenues towards the realization of in-situ photonic neural network for future edge AI, smart cities, automated driving and supercomputer.

“There is a search for more efficient optoelectronic devices to enable photonic neuromorphic computing. The key components are non-volatile analog switch and thresholding device. Here is a perfect proposal of devices with novel materials that can enable them.” commented Professor Thean.

The first author of the published research is NUS graduate student Zefeng Xu, a member of Professor Thean’s laboratory, who is pursuing PhD-MBA double degree concurrently. He believes that “photonics could revolutionize artificial intelligence and motivate a huge emerging market”.

---

---